Fault detection for high-speed train traction system using autoencoder-Fréchet inception distance

Abstract The detection of the process state in the traction system plays a crucial role in ensuring the safe operation of high-speed trains. However, the running data in the traction system exhibits nonlinear and dynamic nature, which leads to the failure of traditional detection methods. To address the above problems, a novel fault detection method based on autoencoder-Fréchet inception distance (AE-FID) is proposed in this paper. Firstly, autoencoder (AE) is used for feature extraction to capture the inherent nonlinear structure in the data. Secondly, a novel fault detection statistic is developed by combining sliding window technology and Fréchet inception distance (FID) to eliminate the influence of data dynamics. Finally, the proposed statistic is utilized to detect faults in the traction system, and the occurrence of faults is measured in terms of structural differences to mitigate the influence of the masking effects. The proposed method is evaluated through a numerical example that simulates dynamic nonlinear systems and the Traction Drive Control System-Fault Injection Benchmark simulation platform of a high-speed train. The experimental results demonstrate the effectiveness of the proposed method in fault detection for traction systems.